1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Generic INET transport hashtables
9 * Authors: Lotsa people, from code originally in tcp
12 #include <linux/module.h>
13 #include <linux/random.h>
14 #include <linux/sched.h>
15 #include <linux/slab.h>
16 #include <linux/wait.h>
17 #include <linux/vmalloc.h>
18 #include <linux/memblock.h>
20 #include <net/addrconf.h>
21 #include <net/inet_connection_sock.h>
22 #include <net/inet_hashtables.h>
23 #if IS_ENABLED(CONFIG_IPV6)
24 #include <net/inet6_hashtables.h>
26 #include <net/secure_seq.h>
29 #include <net/sock_reuseport.h>
31 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
32 const __u16 lport, const __be32 faddr,
35 static u32 inet_ehash_secret __read_mostly;
37 net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
39 return __inet_ehashfn(laddr, lport, faddr, fport,
40 inet_ehash_secret + net_hash_mix(net));
43 /* This function handles inet_sock, but also timewait and request sockets
46 static u32 sk_ehashfn(const struct sock *sk)
48 #if IS_ENABLED(CONFIG_IPV6)
49 if (sk->sk_family == AF_INET6 &&
50 !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
51 return inet6_ehashfn(sock_net(sk),
52 &sk->sk_v6_rcv_saddr, sk->sk_num,
53 &sk->sk_v6_daddr, sk->sk_dport);
55 return inet_ehashfn(sock_net(sk),
56 sk->sk_rcv_saddr, sk->sk_num,
57 sk->sk_daddr, sk->sk_dport);
61 * Allocate and initialize a new local port bind bucket.
62 * The bindhash mutex for snum's hash chain must be held here.
64 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
66 struct inet_bind_hashbucket *head,
67 const unsigned short snum,
70 struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
73 write_pnet(&tb->ib_net, net);
77 tb->fastreuseport = 0;
78 INIT_HLIST_HEAD(&tb->owners);
79 hlist_add_head(&tb->node, &head->chain);
85 * Caller must hold hashbucket lock for this tb with local BH disabled
87 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
89 if (hlist_empty(&tb->owners)) {
90 __hlist_del(&tb->node);
91 kmem_cache_free(cachep, tb);
95 bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
96 unsigned short port, int l3mdev)
98 return net_eq(ib_net(tb), net) && tb->port == port &&
102 static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb,
104 struct inet_bind_hashbucket *head,
105 unsigned short port, int l3mdev,
106 const struct sock *sk)
108 write_pnet(&tb->ib_net, net);
111 #if IS_ENABLED(CONFIG_IPV6)
112 tb->family = sk->sk_family;
113 if (sk->sk_family == AF_INET6)
114 tb->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
117 tb->rcv_saddr = sk->sk_rcv_saddr;
118 INIT_HLIST_HEAD(&tb->owners);
119 hlist_add_head(&tb->node, &head->chain);
122 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
124 struct inet_bind_hashbucket *head,
127 const struct sock *sk)
129 struct inet_bind2_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
132 inet_bind2_bucket_init(tb, net, head, port, l3mdev, sk);
137 /* Caller must hold hashbucket lock for this tb with local BH disabled */
138 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
140 if (hlist_empty(&tb->owners)) {
141 __hlist_del(&tb->node);
142 kmem_cache_free(cachep, tb);
146 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
147 const struct sock *sk)
149 #if IS_ENABLED(CONFIG_IPV6)
150 if (sk->sk_family != tb2->family)
153 if (sk->sk_family == AF_INET6)
154 return ipv6_addr_equal(&tb2->v6_rcv_saddr,
155 &sk->sk_v6_rcv_saddr);
157 return tb2->rcv_saddr == sk->sk_rcv_saddr;
160 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
161 struct inet_bind2_bucket *tb2, unsigned short port)
163 inet_sk(sk)->inet_num = port;
164 sk_add_bind_node(sk, &tb->owners);
165 inet_csk(sk)->icsk_bind_hash = tb;
166 sk_add_bind2_node(sk, &tb2->owners);
167 inet_csk(sk)->icsk_bind2_hash = tb2;
171 * Get rid of any references to a local port held by the given sock.
173 static void __inet_put_port(struct sock *sk)
175 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
176 struct inet_bind_hashbucket *head, *head2;
177 struct net *net = sock_net(sk);
178 struct inet_bind_bucket *tb;
181 bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
182 head = &hashinfo->bhash[bhash];
183 head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
185 spin_lock(&head->lock);
186 tb = inet_csk(sk)->icsk_bind_hash;
187 __sk_del_bind_node(sk);
188 inet_csk(sk)->icsk_bind_hash = NULL;
189 inet_sk(sk)->inet_num = 0;
190 inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
192 spin_lock(&head2->lock);
193 if (inet_csk(sk)->icsk_bind2_hash) {
194 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
196 __sk_del_bind2_node(sk);
197 inet_csk(sk)->icsk_bind2_hash = NULL;
198 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
200 spin_unlock(&head2->lock);
202 spin_unlock(&head->lock);
205 void inet_put_port(struct sock *sk)
211 EXPORT_SYMBOL(inet_put_port);
213 int __inet_inherit_port(const struct sock *sk, struct sock *child)
215 struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
216 unsigned short port = inet_sk(child)->inet_num;
217 struct inet_bind_hashbucket *head, *head2;
218 bool created_inet_bind_bucket = false;
219 struct net *net = sock_net(sk);
220 bool update_fastreuse = false;
221 struct inet_bind2_bucket *tb2;
222 struct inet_bind_bucket *tb;
225 bhash = inet_bhashfn(net, port, table->bhash_size);
226 head = &table->bhash[bhash];
227 head2 = inet_bhashfn_portaddr(table, child, net, port);
229 spin_lock(&head->lock);
230 spin_lock(&head2->lock);
231 tb = inet_csk(sk)->icsk_bind_hash;
232 tb2 = inet_csk(sk)->icsk_bind2_hash;
233 if (unlikely(!tb || !tb2)) {
234 spin_unlock(&head2->lock);
235 spin_unlock(&head->lock);
238 if (tb->port != port) {
239 l3mdev = inet_sk_bound_l3mdev(sk);
241 /* NOTE: using tproxy and redirecting skbs to a proxy
242 * on a different listener port breaks the assumption
243 * that the listener socket's icsk_bind_hash is the same
244 * as that of the child socket. We have to look up or
245 * create a new bind bucket for the child here. */
246 inet_bind_bucket_for_each(tb, &head->chain) {
247 if (inet_bind_bucket_match(tb, net, port, l3mdev))
251 tb = inet_bind_bucket_create(table->bind_bucket_cachep,
252 net, head, port, l3mdev);
254 spin_unlock(&head2->lock);
255 spin_unlock(&head->lock);
258 created_inet_bind_bucket = true;
260 update_fastreuse = true;
263 } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
264 l3mdev = inet_sk_bound_l3mdev(sk);
267 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
269 tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
276 if (update_fastreuse)
277 inet_csk_update_fastreuse(tb, child);
278 inet_bind_hash(child, tb, tb2, port);
279 spin_unlock(&head2->lock);
280 spin_unlock(&head->lock);
285 if (created_inet_bind_bucket)
286 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
287 spin_unlock(&head2->lock);
288 spin_unlock(&head->lock);
291 EXPORT_SYMBOL_GPL(__inet_inherit_port);
293 static struct inet_listen_hashbucket *
294 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
298 #if IS_ENABLED(CONFIG_IPV6)
299 if (sk->sk_family == AF_INET6)
300 hash = ipv6_portaddr_hash(sock_net(sk),
301 &sk->sk_v6_rcv_saddr,
302 inet_sk(sk)->inet_num);
305 hash = ipv4_portaddr_hash(sock_net(sk),
306 inet_sk(sk)->inet_rcv_saddr,
307 inet_sk(sk)->inet_num);
308 return inet_lhash2_bucket(h, hash);
311 static inline int compute_score(struct sock *sk, struct net *net,
312 const unsigned short hnum, const __be32 daddr,
313 const int dif, const int sdif)
317 if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
318 !ipv6_only_sock(sk)) {
319 if (sk->sk_rcv_saddr != daddr)
322 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
324 score = sk->sk_bound_dev_if ? 2 : 1;
326 if (sk->sk_family == PF_INET)
328 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
334 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
335 struct sk_buff *skb, int doff,
336 __be32 saddr, __be16 sport,
337 __be32 daddr, unsigned short hnum)
339 struct sock *reuse_sk = NULL;
342 if (sk->sk_reuseport) {
343 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
344 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
350 * Here are some nice properties to exploit here. The BSD API
351 * does not allow a listening sock to specify the remote port nor the
352 * remote address for the connection. So always assume those are both
353 * wildcarded during the search since they can never be otherwise.
356 /* called with rcu_read_lock() : No refcount taken on the socket */
357 static struct sock *inet_lhash2_lookup(struct net *net,
358 struct inet_listen_hashbucket *ilb2,
359 struct sk_buff *skb, int doff,
360 const __be32 saddr, __be16 sport,
361 const __be32 daddr, const unsigned short hnum,
362 const int dif, const int sdif)
364 struct sock *sk, *result = NULL;
365 struct hlist_nulls_node *node;
366 int score, hiscore = 0;
368 sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
369 score = compute_score(sk, net, hnum, daddr, dif, sdif);
370 if (score > hiscore) {
371 result = lookup_reuseport(net, sk, skb, doff,
372 saddr, sport, daddr, hnum);
384 static inline struct sock *inet_lookup_run_bpf(struct net *net,
385 struct inet_hashinfo *hashinfo,
386 struct sk_buff *skb, int doff,
387 __be32 saddr, __be16 sport,
388 __be32 daddr, u16 hnum, const int dif)
390 struct sock *sk, *reuse_sk;
393 if (hashinfo != net->ipv4.tcp_death_row.hashinfo)
394 return NULL; /* only TCP is supported */
396 no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP, saddr, sport,
397 daddr, hnum, dif, &sk);
398 if (no_reuseport || IS_ERR_OR_NULL(sk))
401 reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
407 struct sock *__inet_lookup_listener(struct net *net,
408 struct inet_hashinfo *hashinfo,
409 struct sk_buff *skb, int doff,
410 const __be32 saddr, __be16 sport,
411 const __be32 daddr, const unsigned short hnum,
412 const int dif, const int sdif)
414 struct inet_listen_hashbucket *ilb2;
415 struct sock *result = NULL;
418 /* Lookup redirect from BPF */
419 if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
420 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
421 saddr, sport, daddr, hnum, dif);
426 hash2 = ipv4_portaddr_hash(net, daddr, hnum);
427 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
429 result = inet_lhash2_lookup(net, ilb2, skb, doff,
430 saddr, sport, daddr, hnum,
435 /* Lookup lhash2 with INADDR_ANY */
436 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
437 ilb2 = inet_lhash2_bucket(hashinfo, hash2);
439 result = inet_lhash2_lookup(net, ilb2, skb, doff,
440 saddr, sport, htonl(INADDR_ANY), hnum,
447 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
449 /* All sockets share common refcount, but have different destructors */
450 void sock_gen_put(struct sock *sk)
452 if (!refcount_dec_and_test(&sk->sk_refcnt))
455 if (sk->sk_state == TCP_TIME_WAIT)
456 inet_twsk_free(inet_twsk(sk));
457 else if (sk->sk_state == TCP_NEW_SYN_RECV)
458 reqsk_free(inet_reqsk(sk));
462 EXPORT_SYMBOL_GPL(sock_gen_put);
464 void sock_edemux(struct sk_buff *skb)
466 sock_gen_put(skb->sk);
468 EXPORT_SYMBOL(sock_edemux);
470 struct sock *__inet_lookup_established(struct net *net,
471 struct inet_hashinfo *hashinfo,
472 const __be32 saddr, const __be16 sport,
473 const __be32 daddr, const u16 hnum,
474 const int dif, const int sdif)
476 INET_ADDR_COOKIE(acookie, saddr, daddr);
477 const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
479 const struct hlist_nulls_node *node;
480 /* Optimize here for direct hit, only listening connections can
481 * have wildcards anyways.
483 unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
484 unsigned int slot = hash & hashinfo->ehash_mask;
485 struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
488 sk_nulls_for_each_rcu(sk, node, &head->chain) {
489 if (sk->sk_hash != hash)
491 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
492 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
494 if (unlikely(!inet_match(net, sk, acookie,
495 ports, dif, sdif))) {
503 * if the nulls value we got at the end of this lookup is
504 * not the expected one, we must restart lookup.
505 * We probably met an item that was moved to another chain.
507 if (get_nulls_value(node) != slot)
514 EXPORT_SYMBOL_GPL(__inet_lookup_established);
516 /* called with local bh disabled */
517 static int __inet_check_established(struct inet_timewait_death_row *death_row,
518 struct sock *sk, __u16 lport,
519 struct inet_timewait_sock **twp)
521 struct inet_hashinfo *hinfo = death_row->hashinfo;
522 struct inet_sock *inet = inet_sk(sk);
523 __be32 daddr = inet->inet_rcv_saddr;
524 __be32 saddr = inet->inet_daddr;
525 int dif = sk->sk_bound_dev_if;
526 struct net *net = sock_net(sk);
527 int sdif = l3mdev_master_ifindex_by_index(net, dif);
528 INET_ADDR_COOKIE(acookie, saddr, daddr);
529 const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
530 unsigned int hash = inet_ehashfn(net, daddr, lport,
531 saddr, inet->inet_dport);
532 struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
533 spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
535 const struct hlist_nulls_node *node;
536 struct inet_timewait_sock *tw = NULL;
540 sk_nulls_for_each(sk2, node, &head->chain) {
541 if (sk2->sk_hash != hash)
544 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
545 if (sk2->sk_state == TCP_TIME_WAIT) {
547 if (twsk_unique(sk, sk2, twp))
554 /* Must record num and sport now. Otherwise we will see
555 * in hash table socket with a funny identity.
557 inet->inet_num = lport;
558 inet->inet_sport = htons(lport);
560 WARN_ON(!sk_unhashed(sk));
561 __sk_nulls_add_node_rcu(sk, &head->chain);
563 sk_nulls_del_node_init_rcu((struct sock *)tw);
564 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
567 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
572 /* Silly. Should hash-dance instead... */
573 inet_twsk_deschedule_put(tw);
579 return -EADDRNOTAVAIL;
582 static u64 inet_sk_port_offset(const struct sock *sk)
584 const struct inet_sock *inet = inet_sk(sk);
586 return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
591 /* Searches for an exsiting socket in the ehash bucket list.
592 * Returns true if found, false otherwise.
594 static bool inet_ehash_lookup_by_sk(struct sock *sk,
595 struct hlist_nulls_head *list)
597 const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
598 const int sdif = sk->sk_bound_dev_if;
599 const int dif = sk->sk_bound_dev_if;
600 const struct hlist_nulls_node *node;
601 struct net *net = sock_net(sk);
604 INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
606 sk_nulls_for_each_rcu(esk, node, list) {
607 if (esk->sk_hash != sk->sk_hash)
609 if (sk->sk_family == AF_INET) {
610 if (unlikely(inet_match(net, esk, acookie,
611 ports, dif, sdif))) {
615 #if IS_ENABLED(CONFIG_IPV6)
616 else if (sk->sk_family == AF_INET6) {
617 if (unlikely(inet6_match(net, esk,
619 &sk->sk_v6_rcv_saddr,
620 ports, dif, sdif))) {
629 /* Insert a socket into ehash, and eventually remove another one
630 * (The another one can be a SYN_RECV or TIMEWAIT)
631 * If an existing socket already exists, socket sk is not inserted,
632 * and sets found_dup_sk parameter to true.
634 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
636 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
637 struct inet_ehash_bucket *head;
638 struct hlist_nulls_head *list;
642 WARN_ON_ONCE(!sk_unhashed(sk));
644 sk->sk_hash = sk_ehashfn(sk);
645 head = inet_ehash_bucket(hashinfo, sk->sk_hash);
647 lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
651 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
652 ret = sk_nulls_del_node_init_rcu(osk);
653 } else if (found_dup_sk) {
654 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
660 __sk_nulls_add_node_rcu(sk, list);
667 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
669 bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
672 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
674 this_cpu_inc(*sk->sk_prot->orphan_count);
675 inet_sk_set_state(sk, TCP_CLOSE);
676 sock_set_flag(sk, SOCK_DEAD);
677 inet_csk_destroy_sock(sk);
681 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
683 static int inet_reuseport_add_sock(struct sock *sk,
684 struct inet_listen_hashbucket *ilb)
686 struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
687 const struct hlist_nulls_node *node;
689 kuid_t uid = sock_i_uid(sk);
691 sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
693 sk2->sk_family == sk->sk_family &&
694 ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
695 sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
696 inet_csk(sk2)->icsk_bind_hash == tb &&
697 sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
698 inet_rcv_saddr_equal(sk, sk2, false))
699 return reuseport_add_sock(sk, sk2,
700 inet_rcv_saddr_any(sk));
703 return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
706 int __inet_hash(struct sock *sk, struct sock *osk)
708 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
709 struct inet_listen_hashbucket *ilb2;
712 if (sk->sk_state != TCP_LISTEN) {
714 inet_ehash_nolisten(sk, osk, NULL);
718 WARN_ON(!sk_unhashed(sk));
719 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
721 spin_lock(&ilb2->lock);
722 if (sk->sk_reuseport) {
723 err = inet_reuseport_add_sock(sk, ilb2);
727 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
728 sk->sk_family == AF_INET6)
729 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
731 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
732 sock_set_flag(sk, SOCK_RCU_FREE);
733 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
735 spin_unlock(&ilb2->lock);
739 EXPORT_SYMBOL(__inet_hash);
741 int inet_hash(struct sock *sk)
745 if (sk->sk_state != TCP_CLOSE)
746 err = __inet_hash(sk, NULL);
750 EXPORT_SYMBOL_GPL(inet_hash);
752 void inet_unhash(struct sock *sk)
754 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
759 if (sk->sk_state == TCP_LISTEN) {
760 struct inet_listen_hashbucket *ilb2;
762 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
763 /* Don't disable bottom halves while acquiring the lock to
764 * avoid circular locking dependency on PREEMPT_RT.
766 spin_lock(&ilb2->lock);
767 if (sk_unhashed(sk)) {
768 spin_unlock(&ilb2->lock);
772 if (rcu_access_pointer(sk->sk_reuseport_cb))
773 reuseport_stop_listen_sock(sk);
775 __sk_nulls_del_node_init_rcu(sk);
776 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
777 spin_unlock(&ilb2->lock);
779 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
782 if (sk_unhashed(sk)) {
783 spin_unlock_bh(lock);
786 __sk_nulls_del_node_init_rcu(sk);
787 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
788 spin_unlock_bh(lock);
791 EXPORT_SYMBOL_GPL(inet_unhash);
793 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
794 const struct net *net, unsigned short port,
795 int l3mdev, const struct sock *sk)
797 #if IS_ENABLED(CONFIG_IPV6)
798 if (sk->sk_family != tb->family)
801 if (sk->sk_family == AF_INET6)
802 return net_eq(ib2_net(tb), net) && tb->port == port &&
803 tb->l3mdev == l3mdev &&
804 ipv6_addr_equal(&tb->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
807 return net_eq(ib2_net(tb), net) && tb->port == port &&
808 tb->l3mdev == l3mdev && tb->rcv_saddr == sk->sk_rcv_saddr;
811 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
812 unsigned short port, int l3mdev, const struct sock *sk)
814 #if IS_ENABLED(CONFIG_IPV6)
815 struct in6_addr addr_any = {};
817 if (sk->sk_family != tb->family)
820 if (sk->sk_family == AF_INET6)
821 return net_eq(ib2_net(tb), net) && tb->port == port &&
822 tb->l3mdev == l3mdev &&
823 ipv6_addr_equal(&tb->v6_rcv_saddr, &addr_any);
826 return net_eq(ib2_net(tb), net) && tb->port == port &&
827 tb->l3mdev == l3mdev && tb->rcv_saddr == 0;
830 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
831 struct inet_bind2_bucket *
832 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
833 unsigned short port, int l3mdev, const struct sock *sk)
835 struct inet_bind2_bucket *bhash2 = NULL;
837 inet_bind_bucket_for_each(bhash2, &head->chain)
838 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
844 struct inet_bind_hashbucket *
845 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
847 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
849 #if IS_ENABLED(CONFIG_IPV6)
850 struct in6_addr addr_any = {};
852 if (sk->sk_family == AF_INET6)
853 hash = ipv6_portaddr_hash(net, &addr_any, port);
856 hash = ipv4_portaddr_hash(net, 0, port);
858 return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
861 int inet_bhash2_update_saddr(struct inet_bind_hashbucket *prev_saddr, struct sock *sk)
863 struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
864 struct inet_bind2_bucket *tb2, *new_tb2;
865 int l3mdev = inet_sk_bound_l3mdev(sk);
866 struct inet_bind_hashbucket *head2;
867 int port = inet_sk(sk)->inet_num;
868 struct net *net = sock_net(sk);
870 /* Allocate a bind2 bucket ahead of time to avoid permanently putting
871 * the bhash2 table in an inconsistent state if a new tb2 bucket
874 new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
878 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
881 spin_lock_bh(&prev_saddr->lock);
882 __sk_del_bind2_node(sk);
883 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep,
884 inet_csk(sk)->icsk_bind2_hash);
885 spin_unlock_bh(&prev_saddr->lock);
888 spin_lock_bh(&head2->lock);
889 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
892 inet_bind2_bucket_init(tb2, net, head2, port, l3mdev, sk);
894 sk_add_bind2_node(sk, &tb2->owners);
895 inet_csk(sk)->icsk_bind2_hash = tb2;
896 spin_unlock_bh(&head2->lock);
899 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
903 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
905 /* RFC 6056 3.3.4. Algorithm 4: Double-Hash Port Selection Algorithm
906 * Note that we use 32bit integers (vs RFC 'short integers')
907 * because 2^16 is not a multiple of num_ephemeral and this
908 * property might be used by clever attacker.
909 * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
910 * attacks were since demonstrated, thus we use 65536 instead to really
911 * give more isolation and privacy, at the expense of 256kB of kernel
914 #define INET_TABLE_PERTURB_SHIFT 16
915 #define INET_TABLE_PERTURB_SIZE (1 << INET_TABLE_PERTURB_SHIFT)
916 static u32 *table_perturb;
918 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
919 struct sock *sk, u64 port_offset,
920 int (*check_established)(struct inet_timewait_death_row *,
921 struct sock *, __u16, struct inet_timewait_sock **))
923 struct inet_hashinfo *hinfo = death_row->hashinfo;
924 struct inet_bind_hashbucket *head, *head2;
925 struct inet_timewait_sock *tw = NULL;
926 int port = inet_sk(sk)->inet_num;
927 struct net *net = sock_net(sk);
928 struct inet_bind2_bucket *tb2;
929 struct inet_bind_bucket *tb;
930 bool tb_created = false;
931 u32 remaining, offset;
932 int ret, i, low, high;
937 head = &hinfo->bhash[inet_bhashfn(net, port,
939 tb = inet_csk(sk)->icsk_bind_hash;
940 spin_lock_bh(&head->lock);
941 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
942 inet_ehash_nolisten(sk, NULL, NULL);
943 spin_unlock_bh(&head->lock);
946 spin_unlock(&head->lock);
947 /* No definite answer... Walk to established hash table */
948 ret = check_established(death_row, sk, port, NULL);
953 l3mdev = inet_sk_bound_l3mdev(sk);
955 inet_get_local_port_range(net, &low, &high);
956 high++; /* [32768, 60999] -> [32768, 61000[ */
957 remaining = high - low;
958 if (likely(remaining > 1))
961 get_random_sleepable_once(table_perturb,
962 INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
963 index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
965 offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
968 /* In first pass we try ports of @low parity.
969 * inet_csk_get_port() does the opposite choice.
974 for (i = 0; i < remaining; i += 2, port += 2) {
975 if (unlikely(port >= high))
977 if (inet_is_local_reserved_port(net, port))
979 head = &hinfo->bhash[inet_bhashfn(net, port,
981 spin_lock_bh(&head->lock);
983 /* Does not bother with rcv_saddr checks, because
984 * the established check is already unique enough.
986 inet_bind_bucket_for_each(tb, &head->chain) {
987 if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
988 if (tb->fastreuse >= 0 ||
989 tb->fastreuseport >= 0)
991 WARN_ON(hlist_empty(&tb->owners));
992 if (!check_established(death_row, sk,
999 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
1000 net, head, port, l3mdev);
1002 spin_unlock_bh(&head->lock);
1007 tb->fastreuseport = -1;
1010 spin_unlock_bh(&head->lock);
1015 if ((offset & 1) && remaining > 1)
1016 goto other_parity_scan;
1018 return -EADDRNOTAVAIL;
1021 /* Find the corresponding tb2 bucket since we need to
1022 * add the socket to the bhash2 table as well
1024 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1025 spin_lock(&head2->lock);
1027 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1029 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1030 head2, port, l3mdev, sk);
1035 /* Here we want to add a little bit of randomness to the next source
1036 * port that will be chosen. We use a max() with a random here so that
1037 * on low contention the randomness is maximal and on high contention
1038 * it may be inexistent.
1040 i = max_t(int, i, (prandom_u32() & 7) * 2);
1041 WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
1043 /* Head lock still held and bh's disabled */
1044 inet_bind_hash(sk, tb, tb2, port);
1046 spin_unlock(&head2->lock);
1048 if (sk_unhashed(sk)) {
1049 inet_sk(sk)->inet_sport = htons(port);
1050 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1053 inet_twsk_bind_unhash(tw, hinfo);
1054 spin_unlock(&head->lock);
1056 inet_twsk_deschedule_put(tw);
1061 spin_unlock(&head2->lock);
1063 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1064 spin_unlock_bh(&head->lock);
1069 * Bind a port for a connect operation and hash it.
1071 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1074 u64 port_offset = 0;
1076 if (!inet_sk(sk)->inet_num)
1077 port_offset = inet_sk_port_offset(sk);
1078 return __inet_hash_connect(death_row, sk, port_offset,
1079 __inet_check_established);
1081 EXPORT_SYMBOL_GPL(inet_hash_connect);
1083 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1087 for (i = 0; i <= h->lhash2_mask; i++) {
1088 spin_lock_init(&h->lhash2[i].lock);
1089 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1090 i + LISTENING_NULLS_BASE);
1094 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1095 unsigned long numentries, int scale,
1096 unsigned long low_limit,
1097 unsigned long high_limit)
1099 h->lhash2 = alloc_large_system_hash(name,
1108 init_hashinfo_lhash2(h);
1110 /* this one is used for source ports of outgoing connections */
1111 table_perturb = alloc_large_system_hash("Table-perturb",
1112 sizeof(*table_perturb),
1113 INET_TABLE_PERTURB_SIZE,
1115 INET_TABLE_PERTURB_SIZE,
1116 INET_TABLE_PERTURB_SIZE);
1119 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1121 h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1125 h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1126 /* INET_LHTABLE_SIZE must be a power of 2 */
1127 BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1129 init_hashinfo_lhash2(h);
1132 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1134 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1136 unsigned int locksz = sizeof(spinlock_t);
1137 unsigned int i, nblocks = 1;
1140 /* allocate 2 cache lines or at least one spinlock per cpu */
1141 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1142 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1144 /* no more locks than number of hash buckets */
1145 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1147 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1148 if (!hashinfo->ehash_locks)
1151 for (i = 0; i < nblocks; i++)
1152 spin_lock_init(&hashinfo->ehash_locks[i]);
1154 hashinfo->ehash_locks_mask = nblocks - 1;
1157 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);
1159 struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
1160 unsigned int ehash_entries)
1162 struct inet_hashinfo *new_hashinfo;
1165 new_hashinfo = kmemdup(hashinfo, sizeof(*hashinfo), GFP_KERNEL);
1169 new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
1170 GFP_KERNEL_ACCOUNT);
1171 if (!new_hashinfo->ehash)
1174 new_hashinfo->ehash_mask = ehash_entries - 1;
1176 if (inet_ehash_locks_alloc(new_hashinfo))
1179 for (i = 0; i < ehash_entries; i++)
1180 INIT_HLIST_NULLS_HEAD(&new_hashinfo->ehash[i].chain, i);
1182 new_hashinfo->pernet = true;
1184 return new_hashinfo;
1187 vfree(new_hashinfo->ehash);
1189 kfree(new_hashinfo);
1193 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_alloc);
1195 void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo)
1197 if (!hashinfo->pernet)
1200 inet_ehash_locks_free(hashinfo);
1201 vfree(hashinfo->ehash);
1204 EXPORT_SYMBOL_GPL(inet_pernet_hashinfo_free);